Multimodal Fake News Classification in Tamil Using Fact-Checked Social Media Content and Cost-Sensitive Learning

The exponential rise in user-generated social media content has contributed significantly to the spread of fake news. Regional language communities are particularly vulnerable because these fabricated narratives often integrate manipulated text, images, and audio, which we call multimodal fake news. Tamil language fake news detection presents unique challenges owing to the scarcity of multimodal datasets and the limited adaptation of state-of-the-art methods. This study proposes a transformer-based framework for multimodal fake news classification. It uses Bidirectional Encoder Representations from Transformers (BERT) because of its proven cross-lingual transfer capabilities and contextual understanding advantages in low-resource settings. We introduce TamilFacts, the first comprehensive multimodal dataset comprising 7,934 fact-checked samples, including text, images and speech, curated from Tamil fact-checking websites. The proposed approach integrates contextual embeddings for text, robust image feature extraction methods enhanced by error analysis, and denoised speech representations to capture multimodal context effectively. The experimental results demonstrate that transformer-based architectures efficiently classify fake news with an optimal feature representation and achieve an F1-score of 85%. Additionally, we integrated cost-sensitive learning to mitigate class imbalance by assigning higher weights to minority classes, thereby improving fairness and achieving an accuracy of 76.08%. This study establishes baseline benchmarks for Tamil multimodal fake news detection and provides methodological insights applicable to other regional Indian languages.